Generative adversarial networks (GANs) are widely used in image generation tasks, yet the generated images are usually lack of texture details. In this paper, we propose a general framework, called Progressively Unfreezing Perceptual GAN (PUPGAN), which can generate images with fine texture details. Particularly, we propose an adaptive perceptual discriminator with a pre-trained perceptual feature extractor, which can efficiently measure the discrepancy between multi-level features of the generated and real images. In addition, we propose a progressively unfreezing scheme for the adaptive perceptual discriminator, which ensures a smooth transfer process from a large scale classification task to a specified image generation task. The qualitative and quantitative experiments with comparison to the classical baselines on three image generation tasks, i.e. single image super-resolution, paired image-to-image translation and unpaired image-to-image translation demonstrate the superiority of PUPGAN over the compared approaches.
URL: http://arxiv.org/abs/2006.10250v1